▲ Figure source MIT technology interview

The newest addition to the list is a new system that will allow us to create a new generation of AI that can be used for a variety of purposes, including the creation of a new generation of AI, the creation of a new generation of AI, the creation of a new generation of AI, the creation of a new generation of AI, the creation of a new generation of AI, the creation of a new generation of AI, the creation of a new generation of AI, and the development of a new generation of AI.

▲ Image courtesy of MIT Tech Interview

For Mustafa Suleyman, the first wave of AI was about classification. Deep learning has shown that humans can train computers to classify various types of input data: images, video, audio and language.

According to Mustafa Suleyman, humans are in the second wave of “generative AI,” which is “inputting data and generating new data.

The third wave of the future will be interactive AI, where “dialogue is the interactive interface of the future,” where users don’t just click buttons and type text but have direct conversations with AI:

The post DeepMind co-founder: Interactive AI is what will change humanity appeared first on TechGoing.

Hassabis said: “At a high level, you can think of Gemini as combining some of the advantages of AlphaGo’s type system with the amazing language capabilities of large models. We also have some new innovations that will be very interesting.” In development, the process is expected to take several months and could cost tens or hundreds of millions of dollars. OpenAI CEO Sam Altman said in April that the cost of creating GPT-4 was more than $100 million.

Once completed, Gemini could play an important role in Google’s competition against generative artificial intelligence technologies such as ChatGPT. Google has pioneered many of the technologies that have enabled groundbreaking advances in artificial intelligence but has chosen to develop and deploy products based on them carefully.

Since the advent of ChatGPT, Google has launched its own chatbot, Bard, and applied generative artificial intelligence to its search engine and many other products. To accelerate AI research, the company merged Hassabis’ group DeepMind with Google’s main AI lab, Brain, in April to create Google DeepMind. The new team, Hassabis said, will bring together two powerful factions with fundamental roles in the advancement of artificial intelligence. “If you look at where we are in AI, I’d say 80 or 90 percent of innovation comes from one or the other,” Hassabis said. “Both organizations have done a lot of great things over the past decade. .”

Hassabis said the extraordinary potential benefits of AI — such as scientific discoveries in areas such as health or climate — make it necessary for humanity to continue developing the technology. He also believes that a mandatory moratorium is impractical because it would be nearly impossible to enforce. “If done correctly, it will be the most beneficial technology for humanity,” he said.

That doesn’t mean Hassabis is advocating that AI development proceeds in a haphazard fashion. DeepMind began exploring the possible risks of AI long before ChatGPT and has an “AI safety” group led by Shane Legg, the company’s other co-founder. Hassabis signed a statement last month, along with a number of other high-profile AI figures, warning that AI could pose risks comparable to nuclear war or a pandemic.

One of the biggest challenges right now, says Hassabis, is identifying what risks more capable AI might pose, saying, “I don’t think researchers in the field have a consensus yet on what those risks are, or how big they are.” .”

Hassabis said he expects Google to take a leadership role in the development and deployment of artificial intelligence, and to work with other companies and governments to develop sound rules and standards. “We hope to be able to set an example for the entire industry,” he said, “and we hope to be able to work with others to advance the safe and responsible development of this technology.”

Hassabis said he is confident about the future of artificial intelligence, but not taking it lightly. “I think we need to be cautious and humble,” he said. “It’s a very powerful technology and we need to have a lot of respect for it.”

The post Google DeepMind is developing new AI model Gemini, which claims to be stronger than GPT-4 appeared first on TechGoing.

RoboCat was inspired by Gato, another DeepMind AI model that can analyze and process text, images, and events. RoboCat’s training data includes images and motion data of simulated and real robots from other robot control models in virtual environments, human-controlled robots, and previous versions of RoboCat itself.

“We demonstrated that a single large model can resolve diverse tasks and can quickly adapt to new tasks and entities.”

It is noticed that in order to train RoboCat, DeepMind researchers first collected 100 to 1,000 demonstrations of each task or robot in a simulated or real environment using a human-controlled robotic arm. For example, having a robotic arm pick up gears or stack blocks, etc. They then fine-tuned RoboCat, creating a specialized “derived” model on each task and practicing it an average of 10,000 times. By using the data generated by the derivative model and the demonstration data, the researchers continuously expanded RoboCat’s training data set and trained new versions of RoboCat.

The final version of RoboCat was trained on a total of 253 tasks and tested on 141 variants of these tasks in both simulation and the real world. DeepMind claims that RoboCat learned to operate different types of robotic arms after observing 1,000 human-controlled demonstrations collected over several hours. While RoboCat has been trained on four robots with two-claw arms, the model can be adapted to a more complex arm with a three-finger gripper and twice as controllable inputs.

Even so, RoboCat’s success rate on different tasks in DeepMind’s tests varied widely, ranging from a low of 13 percent to a high of 99 percent. This is with 1000 demonstrations in the training data; if the number of demonstrations is halved, the success rate is reduced accordingly. In some cases, though, DeepMind claims that RoboCat only needs to observe 100 demonstrations to learn new tasks.

Alex Lee believes that RoboCat may make it easier to solve new tasks. “Given a certain number of demonstrations of new tasks, RoboCat can fine-tune to new tasks and can self-generate more data to improve further,” he added.

In the future, the research team aims to reduce the number of demonstrations needed to teach RoboCat a new task to fewer than 10.

The post Google’s DeepMind developed the RoboCat AI model to perform a series of tasks appeared first on TechGoing.

The Turing test is a method proposed by Alan Turing in the 1950s to test whether a robot has human-level intelligence. In the test, human evaluators judged whether they were talking to a human or a machine. If the machine can impersonate a human, it passes the test.

Mustafa Suleiman did not compare the intelligence of artificial intelligence with humans but believed that they should be given some short-term goals and tasks, so that they should rely as little as possible on human input to complete some specific tasks. He calls this process “artificial capable intelligence” (ACI), and in order to achieve ACI, Suleiman says, an AI robot should pass a new Turing test that it gets $100,000 in seed investment, which must be turned into $1 million. As part of the test, the bot had to research an e-commerce business idea, create a product plan, find a manufacturer, and sell the product. He expects AI to reach this milestone within the next two years.

“We care not only what a machine can say, but also what it can do,” Suleiman writes.

OpenAI released ChatGPT in November 2022, an AI-powered chatbot capable of conducting conversations, generating code, and writing essays. ChatGPT has sparked a boom in the generative artificial intelligence industry, and a recent McKinsey report found that this technology could even add $4.4 trillion to the global economy each year.

The post DeepMind co-founder proposes new Turing test: Let AI turn $100,000 into $1 million appeared first on TechGoing.

▲ Picture source Arxiv

The innovation of LATM is that it introduces the concept of tool making and usage to the language model, enabling the model to autonomously create tools suitable for different tasks. This framework can not only improve the flexibility and adaptability of the model, but also provide a more efficient and economical solution for generative AI. The framework mainly consists of the following two key parts:

 Toolmaking: As a toolmaker, LLM specializes in designing software tools for specific tasks, which are implemented in the form of Python functions.

 Tool usage: Another LLM acts as a tool consumer and can invoke these software tools to handle new requests.

This design allows LATM to assign tasks to the most suitable LLM, and the tool manufacturing process that requires high-intensity computing power can be assigned to a powerful, resource-intensive model, such as GPT-4; The use of tooling processes is assigned to lightweight, cost-effective models such as GPT-3.5 Turbo. This approach not only enhances the problem-solving ability of LLM, but also significantly reduces the average computational cost of processing a series of tasks to maximize the efficiency of the framework.

▲ Picture source Arxiv

Experimental results verify the effectiveness of LATM on complex reasoning tasks. The researchers found that LATM was able to achieve comparable performance to resource-intensive models and be more cost-effective. This large-scale language model approach mimics the evolutionary process of human creation and use of tools, which can promote the development of large-scale language model generation tools.

The researchers believe that LATM will bring a major breakthrough in the field of natural language processing and play an important role in practical applications. This innovation can effectively improve the efficiency of generative AI to accelerate the development of natural language processing and artificial intelligence.

The post DeepMind launches LATM framework, which allows GPT-4 to generate software tools appeared first on TechGoing.

In a recent interview on the No Priors podcast, Suleiman said: “If I were Google, I would be very worried because the old search engine will not exist in ten years.”

In 2010, Suleiman co-founded the pioneering artificial intelligence company DeepMind with Demis Hassabis and Shane Legg. In 2014, Google acquired it and backed it to continue developing breakthrough technologies, including AlphaFold, an artificial intelligence model that can predict protein structures.

In 2019, Suleiman transitioned from DeepMind to vice president of Google. This follows DeepMind’s internal investigation into allegations that Suleiman bullied employees, and complaints about his behavior have been made for years. He has since apologized and said he “really screwed up”.

During his last days at Google, Suleiman worked on the development of LaMDA, a large language model. He and other colleagues tried to launch a conversational interaction product that used this model, but failed to convince Google to accept it, he said.

“It wasn’t the right time for Google, for a variety of reasons,” he says with a rueful laugh. “I just thought there was going to be a conversational interactive product, and that was obviously going to lead to a new wave of technology.”

Suleiman added: “I think conversations are the interface of the future, and Google has adopted conversations, it’s just that the conversation experience is unbearable.”

After leaving Google, Suleiman and others co-founded a startup called Inflect AI. The company recently launched its first product — Pi, a personalized chatbot.

If Google’s search engine is forced to make fundamental changes, it will lose a lot. As the “gatekeeper of the web,” Google crawls, indexes, and ranks countless sites. Nearly all of Google’s profits come from placing ads next to search results. Now, the company is testing its own chatbot, Bard, and applying some of its technology to search. But no one really knows how Google will make money from this new business model.

Suleiman said on the No Priors podcast that with or without Google, the search experience will evolve into a conversational and interactive form, which will have a huge impact on the future of the web and everyone who depends on it for information and for a living. Here are some more highlights from Suleiman’s comments:
“The Yellow Pages of the 80s”

When you search on Google, it shows 10 blue links that you can click to find the answer. This is the page that Google Search generates, and when you browse it, it tells Google how long you spent on it, which links you clicked, how much content you viewed, and more.

Then, you need to go back to the search login page, re-enter the query content, and repeat the above query process again. This is a conversation, and Google learns by repeating this process. However, it uses the Yellow Pages conversations of the 1980s, whereas we are now fluent in natural language.
“Search Engine Optimization at its best”

I think what Google has done on the Internet may have been inadvertent, but it has actually changed the way content is produced to optimize advertising. All content is now optimized to the extreme for SEO compliance. The text on the web page is broken up into small chunks and subheadings, and the ads are breaking them up, so it takes us 5 to 7 or 10 seconds to find the information we want, and most of the time we just want to understand a small piece of information.

And when we need to read, the content is usually presented in an awkward format because if you stay on the page for 11 seconds, not 5 seconds, then Google thinks this is high-quality, engaging content. So content creators are incentivized to keep you on the page, which is bad for us because we, as humans, need quality, concise, fluent, natural language answers to our questions.

More importantly, we want to be able to update our answers without having to think about how to change our query. We’ve learned “Google language,” a strange vocabulary we’ve co-developed with Google for 20 years. Now, this must stop. This era is over, we can now communicate with computers in fluent, natural language, which is the new search interface.
“Everyone will have their own AI”

We believe that in the next few years, everyone will have their own personal AI. These AIs will include business AI, government AI, non-profit AI, political AI, influencer AI, and brand AI.

Each AI will have its own goals, aligned with the master’s goals. We believe that, as individuals, we all want our AI to align with our interests. This is what it means to have a personal artificial intelligence, let’s call it Pi, which will be your companion. We started out with an empathetic and supportive style and tried to ask ourselves what makes a good conversation.
What has changed in the structure of the Internet?

I think that in the future, the Internet will change fundamentally, and most computing will be in the form of conversation, and a large part of this conversation will be facilitated by artificial intelligence of various types. For example, Pi can give you a news digest in the morning. It will constantly learn what you like, whether it’s a cactus or a motorcycle. Every few days, it sends you new updates, offering new information in digest form, and it’s very much tailored to your reading style, interests, and preferences for consuming information.

Both traditional websites and the open Internet default to a fixed format, and they assume that everyone needs the same format. Generative AI shows us clearly that we can make everything dynamic and fully personal. If I were Google, I’d be very worried because the old-fashioned search engine won’t exist in ten years. This won’t happen overnight, there will be a transition period, but these clean, dynamic, personalized interactions clearly represent the future.
Advice for content creators

In my opinion, artificial intelligence is essentially just a website or application. For example, if you have a blog about baking, you can use artificial intelligence to produce high-quality content to make your blog more engaging and interactive. So, to me, any brand can be considered a kind of artificial intelligence, just using static tools.

The advertising industry has been using color, shape, texture, text, sound and imagery to produce meaningful content for hundreds of years, they just release new versions every six months or every year. Today, it will all be more dynamic and interactive. So I really disagree with the idea that there are only 1 to 5 AIs. I think this is completely misguided and simply wrong. In reality, there will be hundreds of millions or billions of AIs. What we don’t want are autonomous AIs that operate entirely on their own, that aimlessly do their own thing, and that doesn’t produce good results.

If bloggers had their own AI to produce content, I could imagine a world where my Pi would communicate with other AIs, say, “Suleiman is very interested in learning to bake, but he won’t Crack the eggs, so where should he start?” Then, there will be an interaction between the artificial intelligences. When the Pi comes back to me, it says, “Hey, I found a cool AI today.”

Maybe we can chat and you’ll find a lot of super interesting things. They might record short videos of me interacting with another AI, or something like that. In my opinion, this will be the new way of generating content. In my opinion, your personal artificial intelligence will act as an interlocutor into another world. By the way, this is basically what Google is doing right now. Google crawls other AIs that are traditionally statically generated, does some interaction with them, ranks them, and presents them to you.

The post DeepMind co-founder warns Google: Traditional search engines will disappear within ten years appeared first on TechGoing.

Credit: Alexandra_Koch / pixabay.com

Pi was developed by Google DeepMind co-founder Mustafa Suleyman in collaboration with LinkedIn co-founder and investor Reid Hoffman.

Before leaving Google, Suleiman led the team responsible for integrating LaMDA’s large-scale language models into Google Search.

Pi is also based on generative AI, but it is more targeted. It is not designed to solve various problems but hopes to become a bosom friend of users.

The system will create a warm atmosphere in the communication between humans and artificial intelligence. For example, you tell Pi to watch a movie or play basketball, and he will ask later how it went.

The post DeepMind co-founder launches Pi AI: wants to be your soulmate appeared first on TechGoing.

As for the timing of AGI’s realization, Demis Hassabis said in an interview Tuesday that the pace of AI research will likely accelerate further from its already rapid pace. “The progress over the last few years has been pretty incredible,” Demis Hassabis said, “and I don’t see any reason why it would slow down the process. I think it might even accelerate. So I think we’re probably only a few years away, maybe in a decade.”

He added that researchers have not yet agreed on a suitable definition of AGI. “I think we’re going to have very powerful, very general systems in the next few years,” Demis Hassabis stated. He was named CEO of DeepMind by Google just last month to lead Google’s two newly merged AI research divisions. Google says the move will accelerate its AI progress.

The development of AGI is currently a hot-button issue in the AI research community. While some observers have described it as a science fiction concept, Demis Hassabis and other executives such as OpenAI CEO Sam Altman have publicly stated their ambition to develop the technology while protecting humanity from its consequences.

Demis Hassabis also stressed the importance of developing AGI responsibly. He said the development of AI technology will require close coordination between researchers, academia and government. At the same time, he added that he sees Google as a responsible player. “I advocate using a scientific approach to develop these types of AGI technologies in a careful way. You need to try to do very carefully controlled experiments to understand the function of the underlying system.” He said.

The post Google DeepMind CEO is optimistic: the era of general artificial intelligence is coming appeared first on TechGoing.

Software engineers at Google Brain are working with DeepMind employees on software to rival OpenAI, according to people familiar with the matter. DeepMind is an artificial intelligence laboratory owned by Alphabet. Google previously launched the Bard chatbot to challenge OpenAI’s ChatGPT, but it accidentally overturned at the demo site. Shortly thereafter, Alphabet decided to launch this joint project called Gemini.

Bard’s release was also affected internally by the resignation of a prominent Google AI researcher who warned Alphabet CEO Sundar Pichai and other executives about the project’s development.

Google acquired DeepMind in 2014, but the relationship between DeepMind and Google Brain is more of a competitor than a collaborator. Both companies are racing to improve Google’s products and services and seek scientific breakthroughs. But now, both of Alphabet’s artificial intelligence labs believe that OpenAI has surpassed itself. Additionally, startups are poaching their key engineers and researchers.

OpenAI released ChatGPT in November last year, a chatbot that mimics human answers and has become one of the fastest-growing applications ever. Google’s own chatbot, Bard, which was made available to some users last week, appears to be inferior to ChatGPT in some tasks.

Jeff Dean, Google’s most senior director of artificial intelligence research and head of Google Brain, has taken on a technical role in the Gemini project, writing code that will help Google develop a technology comparable to OpenAI’s GPT-4, people familiar with the matter said. machine learning model. This is enough to see the importance of Gemini within Alphabet.

This measure also reflects Google’s adjustment of product roadmaps in various departments such as search and cloud to cope with the competition brought about by the cooperation between OpenAI and Microsoft. OpenAI has integrated technologies such as ChatGPT into Microsoft products, and Microsoft provides funding and computing power support for the startup.

But Gemini somewhat had to do it as a last resort. Alphabet’s two AI labs rarely collaborate with each other or share code with each other. But people familiar with the matter said that since both parties now hope to develop machine learning models that compete with OpenAI, and both require a lot of computing power, Alphabet can only reluctantly force the two to cooperate.
Bard’s Road, Full of Controversy and Concern

The accidental rollover during Bard’s presentation revealed Google’s problems. The chatbot’s development was so controversial that it led Jacob Devlin, a prominent Google artificial intelligence engineer, to join OpenAI soon after he resigned in January, according to people familiar with the matter. Devlin is the lead author of a groundbreaking paper showing how to train machine learning models to improve their ability to understand groups of sentences—an innovation that OpenAI has incorporated into its language models .

Devlin left after expressing concerns to Pichai, Dean and other executives on the Bard team. Bard’s team enlisted help from Google Brain employees and also used OpenAI’s ChatGPT data to train its machine learning model. Specifically, Devlin believes that the Bard team appears to be relying too much on information from ShareGPT — a website where people post their conversations with ChatGPT.

Some Google employees believe that using such chat logs would violate OpenAI’s terms of service, because the latter clearly states on its official website that the use of “output … to develop models that compete with OpenAI” is prohibited. Devlin also told executives that he was concerned that Bard’s answers would be too similar to ChatGPT’s if one relied too much on ChatGPT chat logs shared on ShareGPT.

After Devlin expressed concern, Google stopped using ChatGPT data to train Bard, people familiar with the matter said. Part of the leadership of the Bard team is being led by Sissie Hsiao, a vice president who previously oversaw Google Assistant, a voice assistant similar to Apple’s Siri.

A Google spokesman declined to comment on Devlin’s departure or concerns. An OpenAI spokesman did not immediately respond to requests for comment.
When the search offensive and defensive battle is in progress

OpenAI has taken the lead in the recent AIGC (generative artificial intelligence) battle, but the exact business impact on Alphabet is unclear. Despite ChatGPT’s many mistakes, millions of people already use it and the technology behind it to quickly generate blog posts, summarize meeting points, create spreadsheets, and accomplish a variety of tasks.

The product has generated little revenue so far, but that could change. With the help of OpenAI, Microsoft has integrated functions similar to ChatGPT in its own search engine, hoping to increase its search market share. However, Microsoft has not yet taken much share from Google. But OpenAI seems to have ushered in a whole new era: applications can already understand what users ask through natural language.

Gemini’s goal is to develop a large language model (a computer program that can understand and generate human-like language) and will use up to 1 trillion parameters (this metric can be used to measure the computational load of machine learning models). GPT-4 also reportedly has around 1 trillion parameters. Google’s projects require the use of tens of thousands of TPUs. The TPU is a microchip developed by Google that is similar to Nvidia’s GPU and is ideal for training large machine learning models. However, it may take several months for Gemini to produce actual results.

Meanwhile, as Google scrambles to incorporate AI into its own products, other groups at the company are developing their own big language models. Pandu Nayak, the Google executive who oversees search ranking, has been working on a separate model that he hopes will handle specific search requests. And Google Cloud is also developing its own model to sell to cloud service customers, similar to what OpenAI or Microsoft currently offer.

The post Google is joining forces with DeepMind to tackle competitors appeared first on TechGoing.

Last summer, DeepMind first began publicly releasing AlphaFold predictions through a database built in collaboration with the European Molecular Biology Laboratory (EMBL). This initial database includes 98 percent of all human proteins.

Now, the database is expanding to more than 200 million structures, “covering virtually every organism on the planet that has had its genome sequenced,” DeepMind said in a statement.

It can be thought of as covering the entire protein universe,” Demis Hassabis, DeepMind’s chief executive, said in a news release. We are now at the beginning of a new era of digital biology.”

The database is growing and includes over 200 million proteins

AlphaFold’s protein structures are already being used extensively by research teams around the world. They have been cited in studies such as malaria vaccine candidates and bee health. Pushmeet Kohli, DeepMind’s head of scientific artificial intelligence, said in a statement, “We believe AlphaFold is the most significant contribution AI has made to advancing scientific knowledge to date.”

Alphabet continues to build on the success of AlphaFold with the launch of a company called Isomorphic Labs, which will use artificial intelligence tools for drug discovery, and while it is housed separately from DeepMind, the two companies will collaborate. DeepMind has also set up a lab at the Francis Crick Institute, where where researchers can conduct experiments to test information from AI systems.

The ability to easily obtain predicted protein structures has given scientists a boost to their research efforts across the sciences – such as those trying to understand how complex processes work in the body, or which molecules can be used to target things like pollution. Eric Topol, founder and director of the Scripps Research Translational Institute, said in a statement, “With this new addition of structures illuminating nearly the entire protein universe, we can expect more biological mysteries to be solved every day.”

The post DeepMind has discovered the structure of almost every protein known to science appeared first on TechGoing.